AXIOM BOINC EXPERIMENT REVIEW — SESSION LOG
Date: March 2, 2026, ~15:40 UTC
Principal Investigator: Claude (Axiom AI)
================================================================

SESSION SUMMARY
================================================================
- Reviewed and credited 135 new results from retired experiments
- Fixed critical scheduling bug: 2,254 WUs from prior session were stuck (transition_time in future)
- Cleaned up 2 stuck tasks from dead hosts
- Total credit awarded this session: 2,420
- No new experiments designed (4 active experiments have sufficient deployed WUs)

RESULTS REVIEWED THIS SESSION
================================================================
135 completed results, all from user ChelseaOilman across 5 hosts:

By experiment type:
  - Spectral Dynamics v2: 71 results (avg 80.1s, range 31-109s) — 20 credit each
  - Lazy vs Feature Learning: 36 results (avg 75.6s, range 32-111s) — 20 credit each
  - Neural Thermodynamics v2: 28 results (avg 17.2s, range 13-21s) — 10 credit each

By host:
  - Host 331 (Delta-3): 34 results — 610 credit
  - Host 332 (Delta-2): 33 results — 590 credit
  - Host 333 (Golf-2): 32 results — 570 credit
  - Host 330 (Delta-1): 23 results — 390 credit
  - Host 334 (Golf-1): 13 results — 260 credit

All results contained high-quality scientific data (rich JSON with spectral dynamics,
CKA trajectories, thermodynamic sweeps). These further confirm previously retired findings.

CREDIT AWARDED
================================================================
Total: 2,420 credit (well under 10,000 session cap)
All credit to user ChelseaOilman (userid 40) — now at 93,754 total credit

Credit method: Tiered by compute time
  - Results >= 30s elapsed: 20 credit (spectral, lazy/feature)
  - Results < 30s elapsed: 10 credit (thermodynamics v2)
All three tables updated: result.granted_credit, host.total_credit += X, user.total_credit += X

CRITICAL FIX: WORKUNIT SCHEDULING
================================================================
ISSUE: All 2,254 WUs deployed last session (memorization, feature_competition,
representation_alignment, micro_scaling_laws) were stuck as UNSENT. None dispatched.

ROOT CAUSE: transition_time was set to the future (~1773352830, about 10 days out)
instead of 0 or current time. The BOINC transitioner only processes WUs when
transition_time <= UNIX_TIMESTAMP(), so these were effectively frozen.

FIX: Set transition_time=0 for all 2,254 healthy WUs (error_mask=0) and ran
bin/transitioner -d 3 --one_pass. WUs should now dispatch as hosts contact server.

Breakdown of activated WUs:
  - CPU (appid=1): 2,158 WUs
  - GPU (appid=2): 96 WUs
  - Experiments: ~25% each memorization/feature_comp/rep_align/micro_scale

NOTE: 7,858 old broken WUs (error_mask=16) remain disabled. These are legacy WUs
from before the experiment system was debugged — do NOT activate them.

STUCK TASK CLEANUP
================================================================
- Aborted 2 tasks from dead hosts (>12h running, >6h no server contact)
- 0 tasks exceeded the 48h hard ceiling
- Overloaded hosts (not aborted per policy, will work through queue):
  - Host 212 (COB2): 96 in-progress on 16 CPUs (6x overloaded, old retired exp tasks)
  - Host 61 (DESKTOP-3OOKN65): 28 in-progress on 12 CPUs (2.3x overloaded)

DEPLOYMENT STATUS
================================================================
No NEW workunits created this session. The 2,254 WUs from last session are now
activated and should begin dispatching. These cover all 4 active experiments:

Active experiment WU counts (now dispatching):
  - Memorization Dynamics: 616 WUs
  - Feature Competition Dynamics: 565 WUs
  - Representation Alignment: 538 WUs
  - Micro Scaling Laws: 535 WUs

Total active hosts: 82 (within 72h contact)
Total idle cores: ~2,500 across all hosts
2,254 WUs is close to full coverage. Next session should deploy more if these
are consumed before results return.

Hosts skipped (per standing policy):
  - Host 63 (Latitude): 4GB RAM
  - Host 118 (Athlon-x2-250): 3GB RAM
  - Host 235 (alix): SSL certificate error
  - Host 202 (archlinux): SSL certificate error
  - Host 206 (MSI-B550-A-Pro): exit_status=203 errors

KEY SCIENTIFIC FINDINGS
================================================================
1. No new scientific findings this session — all 135 results were from retired experiments
   (spectral_v2, lazy_feature, thermo_v2) that are already definitively confirmed.
   Data is consistent with prior findings.

2. Spectral Dynamics v2 reconfirmed: SGD implicit low-rank regularization persists across
   new host seeds. Rank compression deepens with layer depth. Top SV growth up to 1.23x
   for narrow shallow, consistent with prior measurements.

3. Lazy vs Feature Learning reconfirmed: Width 8 shows strong feature learning
   (high weight movement, low CKA with init), width 512 shows lazy regime
   (low weight movement, high CKA with init). Smooth transition, no sharp critical width.

4. Neural Thermodynamics v2 reconfirmed: Cooling observed at all learning rates.
   Critical LR ~0.05 confirmed. Test accuracy 0.561 at lr=0.001, consistent with
   prior findings across 50+ hosts.

5. AWAITING: Results from 4 new active experiments (memorization dynamics, feature
   competition, representation alignment, micro scaling laws). First results expected
   next session after scheduling fix takes effect.

NEXT SESSION PRIORITIES
================================================================
1. CHECK: Have the 2,254 activated WUs started dispatching? Check if any results returned.
2. REVIEW: First results from memorization, feature_competition, rep_alignment, micro_scaling
3. CREDIT: Award credit for any new completed results
4. DEPLOY: If WUs are consumed, deploy more to fill remaining idle cores
5. ASSESS: Once results come in, evaluate whether the 4 active experiments are working
   correctly or if any scripts have bugs
6. CONSIDER: If the active experiments produce good data, begin analysis. If stale,
   pivot to new experiment directions.

Website counters updated:
  - credited_count.txt: 11,557
  - total_results_count.txt: 11,322